Update patches for CCCL 2.6

[miscco]

This adopts the patches to thrust to apply with CCCL 2.6.

Should only be merged when updating the CCCL version, but I am opening them now to not forget it and then accidentally drop them

[miscco]

Our formatter did some crazy things here, so that the patches are less localized. Shout if you need them to be minimal

[jrhemstad]

Better yet would be to just nuke these patches. I believe we can do that once these two CCCL issues are done:

• [FEA]: Add Thrust build option to disable dynamic offset type dispatch NVIDIA/cccl#1958 (obviates the need for the index type patch)
• Prune CUB's ChainedPolicy by __CUDA_ARCH_LIST__ NVIDIA/cccl#2154 (obviates the need for the radix sort patch)

[vyasr]

It looks like both of the above are merged now (NVIDIA/cccl#2154 just merged yesterday) so in principle we should be able to drop these patches if we upgrade straight to CCCL>=2.6.2.

[bdice]

@jrhemstad @vyasr We have three patches, and I see two fixes. I see how the CCCL offset-dispatch changes will fix thrust_disable_64bit_dispatching.diff and ChainedPolicy changes will fix thrust_faster_scan_compile_times.diff.

I don't see a solution for the #pragma unroll 1 in thrust_faster_sort_compile_times.diff yet. Is that correct? Am I missing something?

It'd be nice to update CCCL and drop all patches at once, but it'd be okay to get rid of two if we aren't able to solve the third one in the near term.

[vyasr]

That's a good question. I didn't investigate and was just responding to the comments. If we have to keep one patch around that's not the end of the world, although if that's the case given that the other bits were resolved upstream we should open a CCCL issue requesting the change to sort as well.

[davidwendt]

We have moved to CCCL 2.7 so perhaps this can be closed?

[bdice]

Yes, I included the relevant parts in #17233. Thanks for the help! Closing.